Federated Learning for Privacy-Preserving Deep Learning in IoT Networks FedProx Optimization and Its Impact on Heterogeneity and Efficiency

<jats:p>The wide implementation of IoT devices in smart cities brings some complex security challenges that challenge the capabilities of traditional IDS methods. In response to the challenges posed by data variability due to the distributed nature of IoT devices, this research proposes a Privacy-Preserving Federated Learning Framework with FedProx as an improvement for optimizing the learning process. In order to achieve accurate identification of anomalies, this model is implemented by merging AutoEncoder-DNNs together with Multi-Graph Attention Networks using Wavelet and Fourier transforms. Testing results from this type of implementation displayed high precision, sensitivity, and specificity (99.50% accuracy) when testing for anomalous behavior. The use of FedProx minimizes the amount of communication needed to train and optimizes the convergence speed for this framework, thus making it suitable for implementing a real-world IoT security solution that preserves individual user privacy.</jats:p>